Abstract - Automated seismic fault interpretation has been an active area of research. Since 2018, Deep learning (DL) based
seismic fault interpretation methods have emerged and shown promising results. However, to date, these
methods have not been reasonably summarised, making it difficult for those involved to make sense of the
current development process. To close this gap, we systematically reviewed the DL-based fault interpretation
literature published between 2012 and 2022, and searched seven digital libraries. Fault interpretation has been
considered an image-processing task using only convolutional neural networks (CNN)-based DL methods, and
most of them have been trained in a supervised manner. U-Net and its variants designed for the image segmentation
task are the most commonly used network structures. A total of 73 seismic datasets were summarised
from the 56 articles included, of which only three field datasets and four synthetic datasets were publicly
available benchmarks. The study reported benefits of using DL, such as its outstanding learning and generalisation
capabilities or predicting faults in a fast, cheap and repeatable manner, which ultimately led to an increase
in the acceptability of these methods and the potential to incorporate them into oil and industry
workflows. However, we identified 12 challenges that hinder its integration into industrial workflows, including
the most discussed lack of sufficient annotated data. We conclude with an in-depth discussion of current research
trends and potential future research directions to promote research on less studied areas and collaboration between
computer scientists and geoscientists.
Earth Science Reviews, 243, 104509. doi: https://doi.org/10.1016/j.earscirev.2023.104509, 2023.